Cross-racial envy and underinvestment in South African partnerships

Transcription

1 NOT FOR PUBLIC RELEASE Cambridge Journal of Economics 2008, 1 of 22 doi: /cje/ben011 Cross-racial envy and underinvestment in South African partnerships Daniel Haile, Abdolkarim Sadrieh and Harrie A. A. Verbon* Trust games are employed to investigate the effect of heterogeneity in income and 5 race on cooperation in South Africa. The amount of socio-economic information available to the subjects about their counterparts is varied. No significant behavioural differences are observed when no such information is provided. However, when the information is available, it significantly affects individual trust behaviour. The low income subjects from both racial groups invest significantly less 10 in partnerships with the high income subjects of the other racial group than in any other partnership. We attribute this behaviour to cross-racial envy, which on aggregate may lead to substantial efficiency losses. Key words: Trust game, Ethnic diversity, Income inequality, Cooperation JEL classifications: C91, J Introduction A recent hypothesis in the literature is that economic performance will be furthered by the absence of economic and ethnic divisions (see, e.g., Alesina et al., 1999; Easterly and Levine, 1997). Knack and Keefer (1997), for example, note that, for countries like Norway, Finland, Sweden, Denmark and Canada, a high degree of homogeneity both in 20 terms of income and ethnicity goes along with a high degree of economic performance. 1 The explanation behind this linkage is that people with a greater set of similar characteristics are more likely to form partnerships and start to cooperate. 2 In this paper, we examine the effect of economic and ethnic division in the case of South African society. This society is extreme in its heterogeneity, with segmentation along both 25 racial and income divides. In particular, before the breakdown of apartheid in 1994, the Manuscript received 11 September 2006; final version received 11 September 2007 Address for correspondence: Department of Economics, Tilburg University, PO Box 90153, 5000 LE Tilburg, The Netherlands; H.A.A.Verbon@uvt.nl * Wageningen University, University of Magdeburg and Tilburg University, respectively. 1 Incidentally, based on the World Social Survey measurement, these five countries also have the highest level of trust. The close correlation between trust, social homogeneity and economic performance can also be derived from the experimental results by Glaeser et al. (2000) who find that when individuals are closer socially, trust and trustworthiness tend to be higher. 2 However, as shown by Collier (2001), democracy may considerably soften the detrimental effects of ethnic diversity on economic growth. Ó The Author Published by Oxford University Press on behalf of the Cambridge Political Economy Society. All rights reserved.

2 2 of 22 D. Haile et al. black population 1. received less than 50% of the national income, but this share had risen to 75% in (Stewart, 2000). However, at the same time intra-racial income inequality had increased. Amongst black households the Gini index increased from 0.49 in 1970 to 0.59 in 2000, while among the whites it moved from 0.43 to 0.49 (Whiteford and 30 van Seventer, 2000). Widening income gaps within ethnic groups can imply low intra-group trust levels, next to the low trust between groups. To study both aspects of the trust problem, we conducted trust game experiments in South Africa, in which the subjects were given information about their opponents race and income characteristics. The trust game, originally 35 developed by Berg, Dickhaut and McCabe (1995), is a two-player game in which the first player, the sender, sends part of his initial endowment to the second player, the receiver. The experimenter triples the sent amount, and the receiver can then decide which part of his total endowment (the tripled transferred amount plus the initial endowment) to return to the sender. The game is called a trust game as the amount the sender transfers to 40 the receiver gives an indication of the sender s trust in the willingness of the receiver to reciprocate. Starting with the original inventors of the game, Berg et al. (1995), the trust game has also widely been dubbed the investment game, as the sender, by sending part of the endowment to the receiver, invests his money, which will generate some positive return if the latter returns a sufficiently large amount. Hence, the amount sent is 45 frequently called the investment. The trust game has emerged as one of the leading experimental instrument for the measurement of the level of cooperativeness in societies (see Camerer, 2003; Glaeser et al., 2000) We employ the trust game to assess the degree to which racial and income disparity may be degrading trust in South Africa. More specifically, the questions that we raise include 50 the following. Is the level of trust and cooperative attitude in South African society generally (disregarding specific race and income disparities) different from what has been reported in the literature on other societies? In what ways does information on the race and the income characteristics of a trading partner determine the level of trust and reciprocity in the multiply segmented South African society? The answer to the question whether 55 racial and income divides prevent subjects from investing in personal relationships is obviously vital to economic and social development, because trust and cooperation across racial and income boundaries will be necessary in order to reduce transaction costs and, thus, to enhance economic exchange and efficiency. A number of theories have been proposed to explain why inequality and ethnic 60 heterogeneity may affect the individual level of trust. Since perceived satisfaction (or happiness) is often positively correlated to the perceived relative income position in a reference group (Clark and Oswald, 1996; McBride, 2001; Solnick and Hemenway, 1998), income inequality can affect individuals well-being (Festinger, 1954; Frank, 1985; Pollak, 1976). Hence, high degrees of inequality can lead to envy and dissatisfaction thus 65 reducing trust and cooperativeness in economic relationships (Alesina and La Ferrara, 2002; Leigh, 2006). Inequality can also imply an urge to invest resources to verify the trustworthiness of the individuals with whom they interact. Zak and Knack (2001), in particular, argue that with increasing inequality the transaction costs attached to interaction with individuals of different income levels can become excessively high. 1 According to the 1996 census 77% of the 40,583,573 people in the country were black and 11% were white, while Indians (3%) and coloured (9%) people made up smaller percentages. 2 South Africa had a Gini index of 0.58 in 1997, which made it the country with the highest inequality after Brazil with a Gini index of 0.63 (World Bank, 1997).

3 Cross-racial envy and underinvestment 3 of Ethnic heterogeneity has also been shown to have a negative effect on economic activity (Alesina and La Ferrara, 2002; Bandiera et al., 2005; Leigh, 2006). A number of different causalities for this effect have been mentioned in the literature (cf. Bandiera et al., 2005). Amongst the most prominent hypotheses are that collective action is more difficult in heterogeneous communities due to increased differences in tastes, in preferences for 75 working together in ethnically mixed teams, and in the configuration of the sharing rules. Furthermore, ethnic heterogeneity may also lead to competitive and growth-reducing rentseeking activities (Alesina and La Ferrara, 2005). These effects have been studied on a local or country level by regressing aggregate measures of inequality and heterogeneity on measures of cooperation such as participation in social groups, or on economic growth 80 rates (cf. La Ferrara, 2002). While empirical studies help to reveal the correlations between trust and the socioeconomic parameters, experimental studies, such as ours, scrutinise the underlying behaviour at the individual level. In fact, several experimental studies have examined the effect of ethnicity on trust and trustworthiness. However, no study so far has attempted to 85 disentangle the ethnicity effects from the income inequality effects. Fershtman and Gneezy (2001) studied Israeli partnerships in which the individuals were informed of their partner s last name. Since last names in Israel are generally indicative of their preimmigration ethnicity, the names may be used to discriminate between partners from different ethnical origins. In fact the study showed a systematic mistrust of eastern males, 90 resulting in a lower efficiency of the partnerships they were involved in. On first sight, this outcome seems to imply a clear case for ethnic discrimination. But a closer look at the income distribution across the ethnic divide reveals that there may have been an effect of income inequality that confounded the results. Since individuals of eastern origin are much more likely to be in the lower income groups (Fershtman and Gneezy, 2001), and 95 since the actual income level of the subjects was not controlled for, it remains unclear whether the discrimination (i.e. the distrust) was towards the eastern males or towards poor males. In a study with US subjects, Eckel and Wilson (2003) found that allowing individuals to observe their partner s picture increases trust and trustworthiness. However, they also 100 found that the pictures could lead to ethnic discrimination, because minority groups (in particular African-Americans) were less likely to be trusted than the majority groups (Caucasian). Again, there was no control for the income distribution effects, which happen to point in the same direction as in the Fershtman and Gneezy (2001) study: the ethnic group that is mistrusted happens to be the ethnic group with the lower average income. 105 In a study with South African subjects, Burns (2003) conducted dictator and trust games with high school students in the greater Cape Town area. 1 To check the effect of race on the propensity to trust, subjects were shown pictures of their partners. In the dictator games, blacks are favoured by non-whites, but not by whites, who show no bias towards any race. In trust games, however, black students are trusted less by all groups, 110 including their own group. Once again, this seems to be clear evidence for the prevalence of 1 Experimental studies on trust in (South) Africa using experiments are rare. Barr (2003) conducted experiments in Zimbabwe in order to detect which factors contribute to the feeling of shared social identities within communities. Carter and Casteillo (2003) examine the level of trust for South African communities in the province of KwaZulu-Natal, investigating the difference between rural or urban communities in the degree of intra-racial trust. In a closely connected study, Haddad and Maluccio (2003) conduct householdlevel research in KwaZulu-Natal. Their results suggest that both local trust (in neighbours and extended family) and income level are important for financial group participation, which is shown to be correlated with economic prosperity.

4 4 of 22 D. Haile et al. racial discrimination, but there is no control for possible confounding effects of income inequality. As in the other two studies, the group that is being mistrusted most also represents the poorest ethnic group, leaving the question open as to whether racial or income discrimination has been detected. 115 Our experimental design allows us to disentangle the two main division lines that exist in heterogeneous societies (i.e. the racial or ethnic and the income divides) by giving subjects information both on the income level and the race of their partner. Surprisingly, we find neither a purely racial nor a purely income-based discrimination effect. Instead, we discover a strong and significant cross-racial envy effect that to our knowledge has not 120 been reported by any study so far. Envy is known to play an important role in social and economic life (Mui, 1995; Schoeck, 1966). An envious person suffers from the knowledge that another person has a greater share of the economic resources. Envious individuals, therefore, may be willing to take actions to decrease other individuals resources, even if such actions come at a strictly positive cost. 125 In the context of our trust game, if low income subjects have feelings of envy towards high income subjects, they will invest or return little when matched with a high income partner. We find no sign of envy between individuals of the same race. What we find is that low income individuals of both races display envy only in partnerships with a high income partner from the other race. This effect can be considered a robust discrimination effect, 130 because behaviour in the baseline treatment without information, which we conducted with a randomised sub-sample of our subjects, neither exhibits significant differences within the subject pool, nor significant differences when compared to the behaviour observed in earlier experiments without information. Next to envy, our experimental design also allows us to discover charity, the behavioural 135 counterpart to envy. While an envious person has incentives to decrease the payoffs of those who are better-off, a charitable person has incentives to increase the payoffs of others, and the more so the lower their income. 1 Hence, charity, in our experiment, entails investing more than one expects to receive back or returning more than the investment. The only charitable behaviour we observe pertains to high income white individuals who 140 are matched to low income black individuals. The observed effects are likely to be a remnant of the history of apartheid and antiapartheid in South Africa. The violent inter-racial tensions seem to have focused the envy across the race barrier, leaving the poor black especially envious of the rich white and the poor white especially envious of the rich black. Surprisingly, we find no pure racial 145 discrimination effect, which hints at a sense of solidarity and acceptance across races amongst individuals of the same economic segments. Furthermore, we even find some charitable behaviour across races, with charity mainly going from the rich white to the less privileged black. We proceed in the following fashion. Section 2 describes the experimental protocol. 150 Section 3 contains the results on the effects of information. In particular it discusses whether discrimination, if it exists, is based on income or on race. Section 4 works out the correlates between survey questions and experimental decisions and thereby analyses the impact of social distance on the propensity to trust. The last section concludes. 1 Our design also allows us to discover other forms of altruism, such as warm-glow altruism, in which giving is independent of the distribution of payoffs. However, we do not find any signs of such behaviour.

5 2. Experimental procedures Cross-racial envy and underinvestment 5 of The subjects played the one-shot trust game. We applied the strategy method where each subject decided how much to transfer both in the role of sender and the role of receiver. 1 Subjects knew at the start of the experiment that they had to play both roles. The role that determined the actual payoff was drawn randomly. Both sender and receiver were endowed with 20 Rand. (At the time of the experiment the exchange rate 160 was Rand.) As senders, subjects were asked to decide how much of their endowment they would like to transfer to a receiver. To decrease decision complexity and paperwork, the sender transfer was restricted to being zero or any even integer smaller or equal to 20. As receivers, subjects were asked to decide how much they would like to transfer back for each of the 11 possible amounts (0, 2,..., 20) that they may 165 have received from a sender. Additionally, each subject was asked to report both the amount expected as a return on their own transfer as a sender and the amount expected as an investment as a receiver. 2 Our treatment variable is the information that subjects received on the characteristics of their counterparts. Upon recruitment, subjects were asked to report their race 3 and 170 to assess whether they perceive their family income to be above or below the average family income in South Africa. We chose to use a perception instrument instead of an exact measurement, because it is more likely that behaviour is linked to the perception of the relative income position than to an exact measure of income differences, which is not readily accessible to most people. Based upon this self-assessment information, we 175 categorised our subjects according to their race (B 5 black; W 5 white) and their income level (L 5 low income, i.e. below average; H 5 high income, i.e. above average). In the information treatment, the information on the race and the income level of the counterpart was given to each subject before any decision was made. The information was given in the same brief and matter-of-fact manner in all information 180 sessions. No specific emphasis was put on any part of the information (see the instructions in Appendix A). In the no information treatment, no such information was given to subjects. The experimental conditions, including the number of observations for all the distinguished subject types and treatments, are summarised in Table A total of 172 subjects participated in the experiment: 94 black and 78 white. Family income was considered to be below average by 112 subjects and above average by Obviously, the population distribution at the universities is not equivalent to the distribution of race and income in the general population of South Africa. But, note that this does not disturb the internal validity of our experimental investigation, because the 190 socio-economic matching that we used in each partnership was predefined and known to the subjects. Hence, the population distribution played absolutely no role in the decisionmaking and the evaluation of the results. 1 We applied the strategy method, as we had to recruit a large part of our subjects on the spot and did not know in advance whether a counterpart for each recruited subject could be found. 2 The instructions to the subjects are contained in the Appendix. 3 In South Africa students are normally asked to indicate their race upon registering for a university or a school. Hence, asking for this information does not have to generate suspicion with the subjects regarding the experimenters intentions. No single subject objected to providing the information on their race. 4 We used an open advertisement recruiting procedure at the two universities visited. Even though the student population can be considered as biased towards the high income groups, the frequency of high income blacks in the population is so low, that we were not able to recruit as many BH subjects as we had initially planned.

6 6 of 22 D. Haile et al. Table 1. Experimental conditions and number of observations in South Africa trust game No information a Information a Total BL BH WL WH Subjects characteristics BL BH WL WH Total a In the information treatment, subjects were given both race and income level information on their counterpart before the transfer decisions were made. No such information was given at any time in the no information treatment. BL, black on low (below average) income; BH, black on high (above average) income; WL, white on low income; WH, white on high income. Note that income categories are based on subject s own self-assessment. The experimental sessions were conducted in October 2003, at the Potchefstroom University 1. (predominantly white) and at the Mafeking University (predominantly black). 195 We conducted the experiment using pen and paper. A post-experimental questionnaire, containing some general questions and some standardised items on equity preferences, was solicited from each subject after the experiment. In particular, subjects had to react to the well-known trust question from the General Social Survey (GSS), used by Knack and Keefer (1997) in their demonstration that trust is conducive to economic growth. In 200 Section 4 we will consider whether the extracted information in the questionnaire contribute to explaining the experimental results. After all sessions were completed, each subject s decision form was linked to that of another subject, respecting the predetermined matching that was recorded on the subject s decision form. 2 Next, the role of the subject in the partnership was determined randomly and with equal probabilities. Finally, the 205 subject s payoff was calculated using the linked decision forms and the subjects were paid in cash. Note that our subjects have an incentive to act according to their self-interest and not according to the perceived interest of the experimenter, because their payoffs are incentive compatible and not flat (as in many surveys and in some psychological experiments). 210 Furthermore, our subjects are not confronted with survey-like questions and receive neither explicit nor subtle hints as to what the experimenter expects to observe. In fact, in the given situation, it is rather difficult to pinpoint the socially desirable behaviour, since the game begins with a fair division (every player receives the same endowment) and trusting the other is not obviously fairness enhancing. Finally, we can verify that our 215 procedures did not create a distorting experimenter demand effect, because the responses generated in our control treatment (without information on the race or the income level of the counterpart) show no significant differences when compared with the responses observed in the original double-blind experiment by Berg et al. (1995). Their double-blind 1 Conducting the experiments was made possible by the hospitality of the Economics Department of Potchefstroom University; in particular, the support given by Professor Wim Naudé was indispensable. 2 Note that the strategy method guarantees that each decision form can be considered as an independent observation, because no interaction has taken place, when subjects make their decisions. Furthermore the strategy method guarantees that we can match unbalanced groups without letting any subject play more than once and without deceiving subjects.

7 Cross-racial envy and underinvestment 7 of 22 procedure safeguards against experimenter demand effects, because it ensures that 220 subjects are fully aware of the fact that the experimenter cannot link observed behaviour to their identities. Another methodological issue concerns the effect the size of the stakes may have on behaviour. The empirical evidence on this issue generally is that high stakes may slightly affect behaviour, but frequently the effects are not even strong enough to be significant [c.f. 225 the overview given by Camerer (2003)]. In particular, quite a number of experiments have been conducted in which the average payoff was close to a monthly or even annual salary of the participants (e.g. Slonim and Roth, 1998). Typically, even in these experiments subjects exhibit very similar patterns of behaviour as in analogue normal stakes experiments with average payoffs close to the opportunity cost of participation, i.e. close 230 to the average expected hourly wage of the group. 3. Results 3.1 Outcomes without race or income information In the original trust game of Berg et al. (1995) subjects were not provided with any socioeconomic information on their counterparts. Nevertheless, it is clear that the subjects had 235 some general notion of the cultural environment in which their partnerships lay. All subjects in that study were university students in the USA. Our subject pool is similar, because all our subjects are university students in South Africa. Clearly, a basic assessment of the general level of trust and trustworthiness exhibited by our subjects is necessary to ensure comparability of our results concerning the main treatment effects (i.e. the effect of 240 socio-economic information on behaviour). Hence, we conducted the no information treatment, which provides us with such a general benchmark, because it elicits behaviour with the same amount of socio-economic information as was given in the original study. In other words, by comparing the results of our no information treatment to the results of Berg et al. (1995), we can examine in which way trust and trustworthiness in South African 245 student communities differs from the US students behaviour, when behaviour in both cases is elicited in the absence of socio-economic discrimination effects. Table 2 presents summary statistics on the original study by Berg et al. (1995) and our two treatments. The table displays the number of independent observations, the initial endowment size, the observed proportion of senders transferring zero, the average 250 investment ratio (i.e. the ratio of the sent amount to the endowment), and the average return ratio (i.e. the ratio of the amount returned to the initial endowment plus received transfer). It is striking how close the values of the three observational variables are when we compare our no information data to the original data. In fact, statistical tests confirm that there is no difference between the behaviour of US and South African subjects in the trust 255 game. 1 Interestingly, this result also stands if we compare the behaviour of the subjects in each of the socio-economic groups of the no information treatment separately. Table 3 reports the investment ratio (both actual and anticipated) as well as the return ratio (both actual and anticipated) for the sub-samples in the no information treatment. 2 Trusting and 1 We use the Mann Whitney U-test to check for location differences between the Berg et al. (1995) data and our no information data. We find significant differences (not even on a 20%, one-tailed level) neither in the portion of senders sending zero, nor in the investment ratio, nor in the return ratio. 2 The BH sub-sample is left out of this analysis, due to the very small number of observations we have.

8 8 of 22 D. Haile et al. Table 2. Comparison of behaviour of subjects between data from Berg et al. (1995) and South Africa Berg et al. (1995) South Africa a No information With information Independent observations Initial endowment size US$10.00 ZAF$20.00 ZAF$20.00 Proportion of first players who sent zero Investment ratio (ratio of investment to endowment) Return ratio (ratio of return to amount available) a In Berg et al. (1995), second players make responses only when senders invest more than zero while in South African experiment subjects continue to play as we used the strategy method. 260 reciprocating behaviour in none of the sub-sample groups of the no information treatment is statistically different from the behaviour of the Berg et al. (1995) subjects. Furthermore, none of the across sub-sample comparisons (e.g. BL versus WL, BL versus WH etc.) reveals a significant difference in behaviour. Thus, despite the fact that South African society is rather heterogeneous, it seems that the general level of trust and 265 trustworthiness (as measured by the trust game) is similar to the level found in more homogenous societies (Camerer, 2003) when discrimination based on race or income information is not possible. 3.2 Aggregate outcomes with race and income information The last column of Table 2 displays the average aggregate outcomes in the information 270 treatment, in which subjects were given socio-economic information on their counterparts. Despite the fact that there are a few more investors sending 0, and that the average aggregate investment is slightly lower in the information than in the no information treatment, we do not observe either strong or significant differences when comparing our treatments to one another, nor when comparing them to the results of Berg et al. (1995). Even a closer 275 look at the relationship between the invested amount and the return ratio does not reveal any behavioural differences on the aggregate level. As displayed by Figure 1, the way the Table 3. Average percentages of available amounts in the no information treatment Socio-economic group BL WL WH Investment ratio Anticipated investment ratio Return ratio Anticipated return ratio BL, black on low (below average) income; WL, white on low income; WH, white on high income. Note that income categories are based on subject s own selfassessment. The BH sub-sample is left out of this analysis, due to the very small number of observations we have

9 Cross-racial envy and underinvestment 9 of Return ratio Return ratio Amounts available in South African Rand No Information Amounts available in South African Rand With Information Fig. 1. Aggregate return ratio at different investment levels. average return ratio increases with the investment level seems to be very similar both with and without socio-economic information. 3.3 Group specific investment decisions with race and income information 280 While the effect of race and income information seems rather small in the aggregate, we can use our detailed data to analyse whether differences exist between the socio-economic groups. Figure 2 shows the average investment ratio exhibited by the subjects in each of the four socio-economic sub-samples (BL, BH, WL and WH) of the information treatment when facing a receiver from their own or from each of the other groups. 1. To facilitate the 285 comparisons, the average investment ratios are shown as deviations from the average investment ratio in the no information treatment (i.e. 0.55). We use the behaviour in the no information treatment as a benchmark, because it cannot be biased by any deliberate discrimination that specifically targets the socio-economic group of the counterpart. Furthermore, since we have established that the behaviour in our no information 290 treatment is indistinguishable from the trust game behaviour observed in completely different cultural settings, the no information benchmark seems to exhibit a certain degree of universality. The leftmost section of Figure 2 displays the average investment ratios chosen by BL subjects in each of their four possible partnerships. Although all four averages lie below the 295 benchmark of the no information treatment, only investment ratios in partnerships with WH subject are, on average, significantly below the benchmark. 2 The third section of Figure 2 shows that WL subjects also tend to invest less when they have socio-economic information on their counterpart than when they do not. Exactly as in the case of the BL subjects, a cross-racial relationship specifies the only partnership in which the investments 300 of the WL subjects are significantly lower than in the no information benchmark. The average investment ratio of the WL subjects in informed partnerships with BH subjects is just above 17%, which is not only dramatically and significantly less than the average 60% 1 Due to a lack of observations we are not able to present any meaningful mutual relationship between BH subjects and WL and BH subjects, respectively. 2 We compare the group s decisions with the overall averages (benchmark case). It would not change our results if comparison were made with group averages. Moreover, we do not find any statistically significant differences in the behaviour of the BL subjects when comparing locations, i.e. Potchefstroom with Mafikeng.

10 10 of 22 D. Haile et al. Investment Investing subjects Black Low (BL) Black High (BH) White Low (WL) White High (WH) a b BL BH WL WH BL WH BL BH WL WH BL BH WL WH Receiving subjects Fig. 2. Average investment ratio. The bars represent the average investment ratio that was exhibited by investors in each socio-economic group (sub-title on top) in partnerships with receivers of each socioeconomic group (indicated on the x-axis). (a) Significantly lower investment ratio in partnerships with WH than in the no information benchmark (at 5%, one-tailed). (b) Significantly lower investment ratio in partnerships with BL than with WH (at 1%, one-tailed). (c) Significantly lower investment ratio in partnerships with BH than in the no information benchmark (at 1%, one-tailed). (d) Significantly lower investment ratio in partnerships with BH than with WL, BL, and WH (at 1%, 1% and 5%, one-tailed, correspondingly). c, d investment ratio of WL subjects in the no information treatment, but also significantly less than any of their average investment ratios in partnerships with any of the other three 305 groups (47.5%, 55.7%, and 52% for BL, WL and WH, respectively). The second and fourth sections of Figure 2 show that, while WH subjects, on average, do not differentiate their investments by socio-economic categories (i.e. there are no significant effects of socio-economic information on investment behaviour), the BH subjects do. The average investment of BH subjects in partnerships with BL subjects is 310 significantly lower than their average investment in partnerships with WH subjects. The average investment ratio of BH in BL partnerships (25%) is also substantially below the BH average investment ratio in the no information benchmark (55%). The lack of statistical significance in this comparison is most probably due to the relatively small number of observations with BH subjects. 315 The fact that subjects differentiate their investments according to the socio-economic characteristics of their counterpart can be based either on a preference for discrimination or on distrust. Distrust towards a certain other group may not be justified, i.e. that group may actually be trustworthy on average, but thought to be the contrary, perhaps due to widespread prejudices in the investor group. But, if the distrust is actually justified, because 320 the average return by partners from the specific socio-economic group is below the return by others, then the low investment may be solely driven by economic incentives and completely free of any preference for discrimination. Figure 3 shows the average return ratio that subjects expected from each of the groups of counterparts. Note that the BL subjects expected almost equal levels of return from all 325 groups. This is important, because it shows that investing very little in partnerships with WH subjects is not due to a low-return expectation, but due to a conscious act of discrimination that is probably based on envy. Note, furthermore, that the observed envy seems to be purely cross-racial, because the BL subjects expectations concerning, and behaviour towards, the rich in their own ethnic group is statistically indistinguishable from 330 the no information benchmark.

11 Anticipated return ratio Black Low (BL) Cross-racial envy and underinvestment 11 of 22 Black High (BH) Investing subjects a White Low (WL) b Returning Subjects c White High (WH) BL BH WL WH BL WH BL BH WL WH BL BH WL WH Fig. 3. Average anticipated return ratio. The bars represent the average return ratio that was expected by investors of each socio-economic group (sub-titles on top) in partnerships with receivers of each socio-economic group (indicated on the x-axis). (a) Significantly lower return ratio expected in partnerships with BL than with WH (at 5%, one-tailed). (b) Significantly lower return ratio expected in partnerships with BH than with WH (at 1%, one-tailed). (c) Significantly lower return ratio expected in partnerships with BL than with BH (at 5%, one-tailed). For the WL subjects things are slightly different. Just as for the BL subjects, WL subjects show cross-racial envy by investing significantly less only in partnerships with BH subjects. The slight difference is that the WL subjects also expect significantly smaller return ratios from the BH subjects than from the WH subjects. Note, however, that this difference is 335 mainly due to the overly optimistic view of the WL subjects on the behaviour of the rich in their own ethnic group. Compared to the benchmark of no information, their return expectations are not significantly different. Summarising, we find discrimination based on cross-racial envy, since both low income type investors (BL and WL) discriminate towards the high income type receivers of the 340 other ethnic group (WH and BH, respectively). In addition to the cross-racial envy effect, we observe that BH subjects discriminate against the poor in their own ethnic group, by investing very little in those partnerships, even though they expect returns from BL subjects to be no different than from any other group. Finally, we also observe mildly charitable behaviour of the WH subjects, who invest substantial amounts in partnerships 345 with BL subjects, even though they expect less to be returned to them by the BL subjects than by any other group. 3.4 Group specific return decisions with race and income information Figure 4 indicates the average return ratio chosen by the subjects of each socio-economic group in their partnerships with subjects of different socio-economic groups. None of the 350 average return ratios is significantly different from the average return ratio observed in the no information treatment. Furthermore, there is only a single case in which the subjects of one group significantly differentiate their return responses on the basis of the socioeconomic characteristics. BL subjects exhibit lower return ratios in partnerships with WH subjects than with WL subjects. This seems to correspond to the cross-racial envy effect 355 that we also observe concerning the investment ratio of the BL subjects. It seems that WH subjects are generally treated somewhat worse by the BL subjects. Note first that this behaviour must be driven by some non-monetary motivation, because the BL subjects

12 12 of 22 D. Haile et al Returning subjects Black Low (BL) Black High (BH) White Low (WL) White High (WH) Return ratio a BL BH WL WH BL WH BL BH WL WH BL BH WL WH Receiving subjects Fig. 4. Average return ratio. The bars represent the average return ratio that was exhibited by receivers in each socio-economic group (sub-title on top) in partnerships with investors of each socioeconomic group (indicated on the x-axis). (a) Significantly lower return ratio in partnerships with WH than with WL (at 5%, one-tailed). invest less in partnerships with WH subjects even though they do not expect lower returns. Note also that the adverse treatment of WH subjects by BL subjects cannot be a purely 360 racial differentiation effect, since we observe a (insignificantly) higher average return ratio of the BL subjects in partnerships with WL subjects than in any partnership with black subjects. Comparing the expressed beliefs of investors and the observed return behaviour, we generally find no significant differences. Hence, expectations and behaviour are pretty 365 well-aligned, perhaps with the notable exception of the white subjects expectations of the black, low-income subjects return behaviour. Although the differences are neither extreme nor significant, it seems that the white subjects, particularly the WH subjects, underestimate the trustworthiness of the BL subjects. Finally, note that the slightly charitable behaviour of WH subjects as investors is 370 echoed in the fact that they, on average, return more to BL subjects than to subjects from any other group. Although the differences in return ratios are not statistically significant, they seem to underline a generally charitable attitude in the group of white, high-income subjects. Figure 5 displays the average investment ratios that were expected in each type of 375 partnership by the receiving subjects in each of the socio-economic groups. As in the previous figures, the average investment ratio expected in the no information treatment is used as a benchmark. In general, the expressed expectations are not significantly distinguishable from the benchmark. Nevertheless, two interesting observation can be made. First, in many cases where an above benchmark investment ratio was expected, the 380 actual investment ratio was below zero and vice versa. In other words, the expectations are not always well aligned with the actual behaviour. Second, the only significant effects observed relate to the expectations of white subjects concerning the investment ratios of BH and WL subjects. All white subjects expect too low investment ratios by BH subjects and too high investment ratios by the WL subjects, when compared with the actual 385 investment behaviour. 3.5 Payoff consequences In the analysis so far, we focused on detecting differential behaviour and expectations. When subjects choices were in conflict with their expectations of monetary income, we

13 Anticipated investment Black Low (BL) Cross-racial envy and underinvestment 13 of 22 Black High (BH) Returning subjects White Low (WL) b BL BH WL WH BL WH BL BH WL WH BL BH WL WH Investing subjects a White High (WH) c Fig. 5. Average anticipated investment ratio. The bars represent the average investment ratio that was expected by receivers in each socio-economic group (sub-title on top) in partnerships with investors of each socio-economic group (indicated on the x-axis). (a) Significantly lower investment ratio expected in partnerships with BH than with WL (at 1%, one-tailed). (b) Significantly lower investment ratio expected in partnerships with WH than with WL (at 5%, one-tailed). (c) Significantly lower investment ratio expected in partnerships with BH than with WL (at 1%, one-tailed). could localise a preference for differential treatment for which a payoff disadvantage was 390 taken into account. In this section we pursue the question how the observed behaviour translates into payoffs for the different subject types. Since we have elicited strategies from our subjects, we can create a more complete picture of the distribution of payoffs by calculating the average payoff each subject would have had if we had matched him/her to every possible counterpart from the corresponding socio-economic group. The top two 395 panels in Figure 6 show the averages of these population payoffs for the investors and the receivers, correspondingly. The lowest panel displays the average income of both roles in each of the socio-economic groups. The no information benchmark is once again provided to allow a visual assessment of the effect of socio-economic information on the payoffs achieved in the game. 400 The general picture that emerges from Figure 6 is that socio-economic information has an adverse effect on the behaviour in partnerships, leading to negative payoff consequences. This holds in particular for BL subjects who are on the losing side in all but one of the possible partnerships. Only BL subjects in partnership with WH subjects earn substantially more than in the no information benchmark. This 405 is mainly due to the fact that WH subjects invest significantly more in partnerships with a BL receiver than in the no information benchmark, although they receive significantly less from the BL subjects. While WH subjects are acting more generously than they would without information, the BL subjects are behaving less generously than they would without information. It is not surprising that the strongest payoff effects 410 of the socio-economic information are present in this specific relationship, in which subjects exhibit preferences for differential treatment that are not based on monetary incentives. 4. Determinants of experimental decisions In this section, we report a series of regression analyses that help to quantify the observed 415 effects of the socio-economic information given to the subjects. Additionally, the

14 14 of 22 D. Haile et al. South African Rand South African Rand Average Payoff as Investors Black Low Black High (BH) White Low (WL) White High (WH) a* a** a*** BL BH WL WH BL WH BL BH WL WH BL BH WL WH Receivers Black Low (BL) Black High (BH) White Low (WL) White High (WH) a** a*** a* a** a*** Average Payoff as Receivers BL BH WL WH BL WH BL BH WL WH BL BH WL WH Investors a*** South African Rand Average Payoff of Players in any Role Black Low (BL) Black High (BH) White Low (WL) White High (WH) a*** a** a** BL BH WL WH BL WH BL BH WL WH BL BH WL WH Pairs Fig. 6. Average payoffs. The bars represent the average payoff of the corresponding players in each socio-economic group (sub-title on top) in partnerships with counterparts of each socio-economic group (indicated on the x-axis). For the WL subjects linked to BH subjects we have no observations on the latter. In this case, therefore, we determined the WL payoffs by drawing from observations from a pilot. Significant at *10% level, one-tailed, **5% level, one-tailed and ***1% level, one-tailed, respectively. a*** regressions are used to check for correlations of the observed behaviour to the questionnaire answers provided by the subjects. We can confirm that low-income subjects send significantly less than other subjects if they are linked to high-income subjects of the other race. Moreover, we show that the questionnaire responses do not provide reliable 420 forecasts on subjects behaviour. First, the regressions explaining the investment ratio are reported and then the regressions explaining the return ratio.

15 Table 4. Double-censored Tobit estimates on the investment ratio 4.1 Investment ratio Constant (0.111)*** (0.109)*** (0.181)*** (0.176) Male (0.071) (0.069) (0.074) (0.063) Low income (0.082)** (0.086) (0.091) (0.078) White (0.078) (0.077) (0.084) (0.070) Income-hetro (0.098) (0.096) (0.097) (0.084) Race-hetro (0.104) (0.101) (0.102) (0.088) Social distance (0.143) (0.194) (0.195) (0.169) Income*Soc (0.184)*** (0.189)*** (0.161)*** Perceptions (0.041) (0.036) Opportunity (0.074) (0.062) Education (0.084) (0.071) Trust index (0.051) (0.044) Expected return ratio (0.161) Expected investment ratio (0.106)*** S (0.031) (0.030) (0.030) (0.025) Log likelihood Wald Test (x 2 ) Observations Numbers in parentheses are the estimated standard errors. **Significant at 5% level; ***significant at 1% level. Cross-racial envy and underinvestment 15 of 22 We used a double-censored Tobit regression with the investment ratio as the dependent variable. The results are summarised in Table 4. The variable names are self-explanatory. 425 (See Appendix Table A.1 for the definitions of variables that are reported.) Column 1 contains the coefficient estimates for the regression in which only the basic characteristics of the sender and the receiver are included. None of the coefficients are significant. In particular, not even the dummy variable social distance, which is equal to one when both the race and the income levels of the matched individuals are different, has 430 significant explanatory power for the observed variations in the chosen investment ratios. This may seem surprising on first sight, but is intuitively clear when we take into consideration that the observed behaviour towards subjects from another race and income group is asymmetric, since the low income subjects hold back investments in partnerships with high income subjects from the other race, but the opposite is not true. In column 2, we 435 introduce a dummy variable called income*soc, which is equal to one when the sender has a low income and is matched with a subject of both a different race and a different income level. As expected, the coefficient of this variable is highly significant, confirming our experimental results. In column 3, questionnaire items measuring subjects perceptions of equity and fairness 440 are included. None of the coefficients are significant. Including trust, measured by means of the GSS trust question as an explanatory variable, does not contribute to explaining the investment ratio. 1 Finally, in column 4, subjects expectation of the amount invested and 1 Incidentally, this confirms results by Glaeser et al. (2000), and subsequently others, who show that trust as measured by the trust game does not correspond to the GSS trust measure.

16 16 of 22 D. Haile et al. returned by their partner is included in the regression. The coefficient on the expected investment ratio is positive and significant, suggesting that, in general, subjects own 445 investment behaviour is driven by what they expect their counterparts to invest, but not on what they return. 4.2 Return ratio The estimated coefficients of the return ratio regressions are summarised in Table 5. Single-censored regression is used as every single subject always retained a substantial part 450 of their partner s investment for themself. Since we used the strategy method, we have more than one observation for each respondent. In particular, each respondent had to state how much to return for each of the 11 possible amounts that could be received from the sender. To take the panel character of this dataset into account, we employed the random effects Tobit model. 455 The return ratio is positively correlated to the amount available to the subject, Q. However, since the return ratio increases as a decreasing rate, there is a negative effect of the square of the amount available. The baseline regression also includes the characteristics of both the investor and responder, as reported in column 1 of Table 5. It seems that males return more, while low income participants return less. Neither income nor race 460 heterogeneity, however, seems to play a role in responder s decision on the return. And again, as in the case of the investments, the symmetric cross-effect of income and race ( social distance ) is not significant, while the asymmetric cross-effect ( income*soc ) has a significantly negative effect on the return ratio, thus conforming the negative effect of cross-cultural envy. 465 Adding the questionnaire responses to the regression (column 3), provides significantly negative, but very small effects for the perceptions and trust index variables, which indicate whether the subject perceives inequality as having been generated fairly and, correspondingly, whether the subject generally trusts others. However, since both coefficients switch to positive as soon as we add the expectation on investment and return 470 (in column 4), it seems that the actual effect of these variables cannot be measured very precisely. However, there is a significantly positive effect between the expected return from the partner and one s own return ratio. 5. Concluding remarks We report the results of a series of experimental trust games conducted in South Africa 475 focusing on the effects of racial and income inequality on cooperation in partnerships. We vary the amount of socio-economic information available to the subjects about their counterparts in order to assess the effect of inequality between the partners. In the control treatment, in which no such information is provided, we observe no significant differences in the behaviour across races and income groups. In fact, despite the extreme heterogeneity 480 of South African society, we find that the general level of trust and trustworthiness observed in the control treatment is very similar to that which is observed in more homogenous societies. When socio-economic information on their counterparts is provided to the subjects, individual trust behaviour is affected significantly. While we neither observe simple racial 485 nor simple income-based discrimination, we do observe that the low income subjects from both racial groups invest significantly less in partnerships with the high income subjects of